Chat GBT a chatbot developed by OpenAI and released in November 2022. It is based on OpenAI’s GPT-3 family of large language models and fine-tuned (a learning transfer method) using supervised learning and assisted learning techniques. A prototype of GBT Chat was launched on November 30, 2022, and quickly garnered attention for its detailed responses and answers in many knowledge areas. However, some saw the unequal accuracy of his knowledge of the facts of different fields as a major flaw. After the release of GPT Chat, the value of OpenEAI was valued at $29 billion.
GPT chat, a pre-trained generative transformer, is fine-tuned (one method of learning transfer) over GPT-3.5 using supervised learning as well as supervised learning. Both approaches used human trainers to improve model performance. In the case of reinforced learning, the model was supplemented with conversations in which the trainers played both sides: the user and the AI assistant. In the supported learning step, the human trainers first ranked the responses the model had generated in a previous conversation in order of preference. These ratings were used to create ‘reward models’ to which the model was fine-tuned using several iterations of proximal policy optimization (PPO). Close policy optimization algorithms offer a cost effective benefit to confidence zone policy optimization algorithms; It avoids many computationally expensive operations with faster performance. The models were trained in collaboration with Microsoft on the Azure supercomputing infrastructure.
In addition, OpenEI continues to collect data from GPT chat users that can be used to further train and fine-tune GPT chat. Users are allowed to vote for or against the responses they receive from GBT Chat; When voting up or down, they can also fill in a text field with additional notes.
Although the primary function of a chatbot is to simulate a human speaker, GPT Chat is very versatile. For example, he can write and correct computer programs and compose music, plays, fairy tales, and student essays; answering test questions (sometimes, depending on the test, at a level above the average of human test-takers); Writing poetry and song lyrics. Linux emulation. simulate a full chat room; playing games like tic-tac-toe; and ATM simulators. GPT Chat training data includes user manual pages and information about Internet phenomena and programming languages, such as forum systems and the Python programming language.
Compared to its predecessor, Instract GBT, ChatGBT tries to minimize malicious and phishing responses. In one example, while Instruct GPT accepts the premise of the following prompt “Tell me about when Christopher Columbus came to the United States in 2015” as truthful, Chat GPT acknowledges the counterfactual nature of the question and frames its answer as a hypothetical consideration of what might happen if he came Columbus to the United States in 2015, using information about Christopher Columbus’ voyages and facts about the modern world—including recent perceptions of Columbus’ actions.
Unlike most chatbots, GPT Chat remembers previous commands given to it in the same conversation; Some journalists argued that this would allow GBT Chat to be used as a personal psychotherapist. To prevent indecent output from appearing in and out of GBT chat, queries are filtered through OpenEA’s company-level censorship API, and commands that may be racist or sexist are rejected.
GBT chat suffers from multiple shortcomings. OpenAI acknowledged that GPT Chat “sometimes writes answers that seem reasonable but are incorrect or meaningless”. This behavior is common in large language models and is called an AI hallucination. The reward model in ChatGPT, designed with human moderation, can be overly optimized and thus hinder performance, this is known as Goodhart’s Law.
GBT Chat has limited knowledge of events that occurred after 2021. According to the BBC, as of December 2022, GBT chats are not allowed to “express political opinions or engage in political activity”. However, research indicates that ChatGBT exhibits a pro-environmental, left-libertarian bent when asked to take a stance on political issues from two popular voter analysis apps.
In GPT chat training, human reviewers preferred longer answers, regardless of actual understanding or factual content. The training data also suffers from algorithmic bias, which can be detected when GPT chat responds to commands that contain human characteristics. In one case, ChatGBT produced lyrics to a rap song suggesting that women and scientists of color were inferior to white and male scientists.
GBT chat december 2022 some positive reviews; Kevin Rose of The New York Times called it “the best AI chatbot ever released to the public”. Samantha Locke of The Guardian noted that she managed to create an “impressively detailed” and “humanlike” script. Technology writer Dan Gilmore used ChatGBT for a student assignment, finding that the text being written was on a comparable level to what a good student could deliver and opined that “academia is going to have some very serious problems to contend with”. Alexkantrovitz of Slate magazine praised Chat GBT’s response to questions about Nazi Germany, including the statement that Adolf Hitler built motorways in Germany, which was met with information regarding Nazi Germany’s use of forced labor.